Generic shape learning and recognition

Abstract

We address the problem of generic shape recognition, in which exact models are not available. We propose an original approach, in which learning and recognition are intimately linked, as recognition is based on previous observation.

The input to our system is in the form of segmented descriptions of objects in terms of parts. In 2-D, the shape is incrementally decomposed into parts suggested by curvature sign changes, and for each part an axial description is derived from both local and global information. The parts are organized into a connection hierarchy. For 3-D objects, we intend to use segmented tridimensional descriptions, the parts being modeled by generalized cylinders. In this case, the connection graph is not necessarily a hierarchy, but can still be used with our algorithms.

The part description obtained at this point is still too detailed and fine grained in order to easily categorize and compare shapes. Hence, we use a simplified description of parts, capturing part local geometry and connection with the superpart information. The local geometry parameters are qualitative and symbolic, and are quasi-invariants under projection and viewpoint change. Both types of parameters take discrete values derived from the available fine description. The connection parameters are normalized to be scale-independent. These simplified part descriptions are organized into a connection hierarchy as provided by the original decomposition. The parameters are chosen to ensure that the information carried in these descriptions is sufficient to perform shape recognition.

Actual shape descriptions are stored in a data-base, from which they must be efficiently and specifically retrieved when a new shape is proposed for recognition. We define a hierarchical indexing system based on the structure of the descriptions and the local description of parts. This mechanism allows for dynamic updating of the data-base with a minimum computing cost.

When a shape is submitted for recognition, the data-base is searched for the closest known shapes. A partial match, based on the connection structure and the aggregation of dissimilarities between parts, is computed incrementally level by level between the shape and the possible candidates. The combination of the incremental process with the hierarchical indexing makes the number of shapes processed at each step decrease rapidly, therefore dramatically reducing the average complexity of the retrieval. The selected retrieved shape(s) are used to give a classification for the submitted shape.

This approach to recognition is influenced by the Case-Based Reasoning (CBR) paradigm, which embeds all the characteristics required to meet our goals, such as the ability to process noisy, incomplete and new data. It also provides an interesting framework for higher-level intelligent processing (e.g. justified interpretation, automatic learning).

We describe our implementation for 2-D shapes recognition and present results. The current implementation should also work on 3-D descriptions as described above, with minor changes. We also intend to use this system as the core of a higher-level vision-based reasoning system.

This research was supported in part by the Advanced Research Projects Agency of the Department of Defense and was monitored by the Air Force Office of Scientific Research under Contract No. F49620-90-C-0078 and/or Grant No F49620-93-1-0620. The United States Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright notation hereon.

Sponsored for this research by NOESIS S.A., Immeuble Ariane, Domaine Technologique de Saclay, 4 rue René Razel - Saclay, F-91892 Orsay Cedex, FRANCE.